Improved Prosody Module in a Text-to-Speech System

Improved prosody module in a Text-to-Speech system Wen-Wei Liao and Jia-Lin Shen Research Center, Delta Electronics, Inc. wei.liau@delta.com.tw, lynn.shen@delta.com.tw Abstract The newly-developed prosody module of our text-to-speech (TTS) system is described in the paper. We present two main works on it’s establishment and improvement. On the basis of potential factors influencing prosody parameters, inclusive of duration, pitch and intensity, the prosody model is built as groundwork of this module which is superior to the former rule-based one in generation of natural prosody. In addition, due to the current model’s flaw in prediction of the pitch contour, we further employ an technique named “Soft Template Mark-up Language“(STEM-ML) to improve the smoothness of intonation which has the crucial influence on the naturalness of synthetic speech. Results of the evaluation indicate that the new prosody model is precise enough to predict reliable prosody parameters’ values and with the STEM-ML technique, the prosody module can further yield 14.75% reduction in the root mean square (RMS) error of the predicted pitch contour. 1. Introduction In consideration of severe limitation in the resource afforded by some applications in need of speech response, we choose to develop one storage-saving TTS system which has functioned successfully in our spoken dialogue system. Accordingly, the acoustic inventory used in our system is simply composed of about four hundred base syllable units whose duration and pitch contour will be modified with the algorithm called Pitch-Synchronous Overlap-Add (PSOLA) [1][12] in the synthesizing phrase. In order to produce natural-sounding synthetic speech, the generation of prosody plays a key role and is a difficult issue yet. Outperforming rule-based method [13][14] which was employed in our system previously, the newly-built statistical model based on sum-of-products approach with key factors affecting prosody [7][8][9][10][11] can predict more accurate values of prosody parameters. And in general, the intonation which is characterized by the pitch contour seems more crucial to the naturalness and intelligibility of synthesized speech in comparison with other prosody elements such as duration, intensity etc [6]. Nevertheless, the pitch contour generated by our current prosody model is still short of smoothness. As a result, we further concentrate our work on this problem. Based on the F0 (fundamental frequency) mean value predicted by the current prosody model, an technique named STEM-ML [2][3][4][5] is adopted to overcome this shortcoming. In the evaluation phrase, we prove that this technique can help to reduce the difference between the predicted and observed pitch contours, which means that a more natural intonation is achieved. The paper is organized as follows. In the chapter 2, we present the prosody modeling in our system, The chapter 3 reports STEM-ML technique and the result of implementation. The conclusion is described in the chapter 4. 2. Prosody modeling In general, prosody mainly consists of duration, pitch, intensity of the spoken unit which is one syllable in terms of Mandarin. Besides, the break between units is one of it’s important elements as well. Therefore, one utterance’s prosody can be regarded as the elaborate composition of these four …

Content truncated to 500 words